PROJECTS

A highlights of what I've designed and built

Antara — Satellite Temporal Super-Resolution via Optical-Flow Frame Interpolation

A capstone ML project targeting the ISRO problem statement "Fill in the Frames Seamlessly," generating a synthetic intermediate frame between two real consecutive GOES-16 geostationary satellite scans to increase effective temporal resolution for tracking fast, non-linear phenomena like cyclones. Built a classical Farneback optical-flow baseline, then fine-tuned a pretrained deep frame-interpolation model (FILM) on real satellite radiance triplets pulled self-supervised from NOAA GOES-16 (no manual labeling — the held-out middle frame is its own ground truth). Curated a principled cyclone-vs-calm evaluation split using IBTrACS storm track data and custom lat/lon-to-satellite-pixel geolocation, showing the learned model degrades far less than classical interpolation under fast cyclone dynamics (headline PSNR/SSIM/LPIPS result). Also ran patch-size and fine-tuning-data-volume ablations, built a self-contained before/after slider demo, and extended the model to continuous-time multi-frame interpolation beyond the trained midpoint.

PythonPyTorchOpenCVFILMOptical FlowFrame InterpolationComputer VisionNumPyxarrayscikit-imageLPIPSGOES-16IBTrACSRemote SensingAWS S3Self-Supervised LearningMatplotlibpytest

Kizuna — Peer-to-Peer Encrypted Chat & Call

Building a decentralized, SSH-inspired encrypted messenger where two peers connect directly by IP and authenticate each other using pinned Ed25519 identity keys and fingerprints, with no central server ever seeing message content. Secure sessions are negotiated with the Noise Protocol Framework (X25519 key exchange, ChaCha20-Poly1305 AEAD) over a custom framed TCP protocol with heartbeats, reconnects, and an explicit connection state machine. Messages, contacts, and files are stored only on the local machine in SQLite. Structured as a Cargo workspace of focused crates (crypto, net, storage, core, CLI) with a Ratatui terminal UI as the first interface; NAT traversal (STUN/TURN/ICE) and encrypted voice calling are planned for later phases.

RustTokioNoise ProtocolsnowEd25519X25519ChaCha20-Poly1305SQLitesqlxRatatuiclapP2P NetworkingNAT TraversalSTUNTURNICEWebRTCCaddyTauri

CodeSync — Real-Time Collaborative Editor

Built a real-time collaborative development workspace where teams can edit code simultaneously with Operational Transform, live presence, and synchronized cursors over WebSockets. Integrated Groq-powered Llama 3.3 70B with token-by-token AI response streaming for every connected user. Added GitHub workspace integration for importing repositories, editing files, committing changes, and pushing directly to GitHub. Supports secure multi-language code execution inside short-lived Docker containers, with persistent sessions, document history, and JWT authentication backed by PostgreSQL.

RustAxumTokioWebSocketsPostgreSQLsqlxJWTNext.jsMonaco EditorZustandDockerGroqCaddyLlama 3.3 70B

Crime Detection System

Crime Detection System is an end-to-end platform that uses pretrained deep-learning models to analyze live and recorded video streams — including secure mobile-camera live streams from smartphones (via IP Webcam Application) with low-latency viewing. It can automatically detect suspicious or criminal activity, generate real-time alerts, log events for review and retraining, and present results through a Next.js frontend and Django backend. Architecture: - Ingest: Video streams ingested by backend streaming service. - Infer: Frames forwarded to ML inference module (local or GPU-enabled). - Persist: Detections logged in backend and optionally stored for retraining. - UI: Frontend consumes API for live view, alerts, and analytics. Models & Data: - Trained Models: BaseModel.h5, CustomisedCNNModel.h5, VGGNet1.h5, CustomCNN.h5. - Data: Collected data and stored in repository.

Django REST FrameworkTensorFlow/KerasWebSocketsPostgreSQL

Dampe – AI Based Cloud Storage

Dampe is a modern, intelligent cloud storage platform that combines secure file storage with AI-powered analysis. Built with cutting-edge technologies, it offers seamless file management, smart content analysis, and universal accessibility through your Gmail account. An AI-powered cloud storage platform built with a scalable full-stack architecture using secure OAuth authentication, Better Auth, and AWS S3-compatible Cloudflare R2 storage. The system integrates AI-driven document analysis, smart search, real-time file management, and a modern responsive dashboard with optimized APIs, cloud-based storage handling, and smooth animated UI/UX.

Next.jsCloudflare R2BucketClerk

Healthify - Medicine Recommendation System

A web app that uses Pandas, Sklearn and ML Techniques for training Model, HTML, CSS for frontend, Flask for backend. – Implemented and evaluated multiple classification models (e.g., Support Vector Classifier, Random Forest, Gradient Boosting) to predict diseases with high accuracy of 85 %. – Implemented typo handling (FuzzyWuzzy) & user authentication (Flask-Login). – Provided treatment recommendations (medications, workouts, diet)..

Machine Learning Flask PythonPandasSklearnFuzzyWuzzyFlask-Login

Newzy-World - AI-Powered News Aggregator

Newzy World is a modern news aggregation and summarization platform powered by AI. Built with Next.js, TypeScript, and Tailwind CSS, it delivers personalized news summaries, real-time updates, and local information in an elegant, user-friendly interface. Features - AI-Powered News: Smart summarization and categorization - Authentication: Secure Google and GitHub OAuth integration - Personalization: User-specific bookmarks and preferences - Weather & Maps: Real-time local information - Modern UI: Responsive design with dark/light mode - Protected Routes: Secure access to user-specific features

Next.jsTypeScriptTailwind CSSAISummarization

Let's Connect

Interested in collaborating or have any questions? Feel free to reach out!

Contact Me